Hybrid MPI/Pthreads parallelization of the RAxML phylogenetics code

Abstract-A hybrid MPI/Pthreads parallelization was implemented in the RAxML phylogenetics code. New MPI code was added to the existing Pthreads production code to exploit parallelism at two algorithmic levels simultaneously: coarse-grained with MPI and fine-grained with Pthreads. This hybrid, multi-grained approach is well suited for current high-performance computers, which typically are clusters of multicore, shared-memory nodes. The hybrid version of RAxML is especially useful for a comprehensive phylogenetic analysis, i.e., execution of many rapid bootstraps followed by a full maximum likelihood search. Multiple multi-core nodes can be used in a single run to speed up the computation and, hence, reduce the turnaround time. The hybrid code also allows more efficient utilization of a given number of processor cores. Moreover, it often returns a better solution than the stand-alone Pthreads code, because additional maximum likelihood searches are conducted in parallel using MPI. The comprehensive analysis algorithm involves four stages, in which coarse-grained parallelism continually decreases from stage to stage. The first three stages speed up well with MPI, while the last stage speeds up only with Pthreads. This leads to a tradeoff in effectiveness between MPI and Pthreads parallelization. The useful number of MPI processes increases with the number of bootstraps performed, but typically is limited to 10 or 20 by the parameters of the algorithm. The optimal number of Pthreads increases with the number of distinct patterns in the columns of the multiple sequence alignment, but is limited to the number of cores per node of the computer being used. For a benchmark problem with 218 taxa, 1,846 patterns, and 100 bootstraps run on the Dash computer at SDSC, the speedup of the hybrid code on 10 nodes (80 cores) was 6.5 compared to the Pthreads-only code on one node (8 cores) and 35 compared to the serial code. This run used 10 MPI processes with 8 Pthreads each. For another problem with 125 taxa, 19,436 patterns, and 100 bootstraps, the speedup on the Triton PDAF computer at SDSC was 38 on two nodes (64 cores) compared to the serial code. This run used 2 MPI processes with 32 Pthreads each.